In an effort to see how effective this cache is in production during this
alpha phase, we'll print out the stats on the document store every 60
seconds.
The implementation of object-size approximation which is used for cache
eviction purposes in the `InMemoryLRUCache` implementation (via `lru-cache`)
was a short-term location for extensible logic which is better located
within `ApolloServerBase`.
This is particularly important since future logic may necessitate knowing or
understanding the current size (roughly, memory usage) of the in-memory
storage. Effective immediately, this adds support for providing a `dispose`
function which is called when an object is purged from the cache to make
room for another.
While the implementation of the `documentStore` is currently simple enough
to never throw (it is in-memory), it makes some sense to guard against future
extended functionality where an exception might be raised.
Since storing this object in a distributed memory store isn't currently
feasible, I'm not sure what such an exception would be right now, but I
don't mind being proactive!
Ref: https://github.com/apollographql/apollo-server/pull/2111/files#r247618501
Without this change, the `document` property was not set on the
`requestContext` for consumption by request pipeline plugins.
To further guard against this oversight, I've removed the extra `document`
variable which was being used as scoped state for the document and switched to
directly using (and assigning to) the `requestContext.document`.
Nice catch, @glasser!
Ref: https://github.com/apollographql/apollo-server/pull/2111/files#r247617469
* refactor: switch `json-stable-stringify` to `fast-json-stable-stringify`
* chore: drop `@types/json-stable-stringify`
* Update CHANGELOG.md for #2065.
This also re-enables the Expiration tests for this library, though I think I
will continue to leave the testsuite decomposed into more granular methods
for readability and future growth.
Note that this doesn't implement the same default `300` second TTL as the
other Memcache and Redis implementations currently do (a very debatable
default we should re-consider in the future).
While it is certainly our current implementation, I'm not sure that the true
spirit of an in-memory key-value store is to only map strings to other strings.
While enforcing `<string, string>` might be necessary for some distributed
cache stores, it seems we shouldn't enforce it for `InMemoryLRUCache`.
By changing the `KeyValueCache` to default to `string` but allow other
options, we can allow the user to decide exactly how the RHS of this store
should be typed.
This does require a bit more flexible implementation of the `length`
calculator which is used for cache ejection, but that implementation will
become immediately useful when we start storing parsed ASTs in this KV store.
